Study shows LLC correlates with neural network compressibility.
problem Evaluating limits of neural network compression.
method Extended minimum description length principle using singular learning theory.
result Complexity estimates based on LLC are linearly correlated with compressibility.
New compression theory justifies model pruning for neural networks.
problem Improving neural network performance with reduced model size.
method Information-theoretic rate-distortion theory applied to NN compression.
result Pruning improves model performance on CIFAR-10 and ImageNet datasets.
New theory controls compression change probability without prior knowledge.
problem Controlling compression change probability without prior knowledge.
method New theory on compression function and statistical risk.
result Cardinality of compressed set is a consistent estimator of probability of change of compression.
Neural NCD reveals LLMs don't compress well for classification.
problem The disconnect between compression and classification in neural networks.
method Developed Neural NCD to compare LLMs to classic algorithms, finding classification accuracy not correlated with compression rate.
result Classification accuracy is not predictable by compression rate alone, challenging current understanding.
One of the biggest issues in deep learning theory is the generalization ability of networks with huge model size. The classical learning theory suggests that overparameterized models cause overfitting. However, practically used large deep models avoid overfitting, which is not well explained by the classical approaches…
Study on reducing forgetting in neural networks using compression theory.
problem Catastrophic forgetting in neural networks.
method Defined forgetting as increased description lengths, compared variational posterior approaches to prequential coding methods.
result Proposed a new continual learning method combining ML plug-in and Bayesian mixture codes.
To improve how neural networks function it is crucial to understand their learning process. The information bottleneck theory of deep learning proposes that neural networks achieve good generalization by compressing their representations to disregard information that is not relevant to the task. However, empirical evid…
The enormous size of modern deep neural networks makes it challenging to deploy those models in memory and communication limited scenarios. Thus, compressing a trained model without a significant loss in performance has become an increasingly important task. Tremendous advances has been made recently, where the main te…
New method connects compression bodies through cone manifolds.
problem Understanding how compression bodies can be transformed.
method Using cone manifold holonomy groups and standard CAT(0) space techniques. result Realized all edges in compression body graph through paths.
Lossless compression of deep neural networks using NTK and RMT.
problem Compressing large-scale deep neural networks for low-power devices.
method High-dimensional neural tangent kernel approach.
result Asymptotic spectral equivalence between NTK matrices of wide DNNs enables lossless compression.
LEAD algorithm speeds up decentralized optimization with compression.
problem Slow convergence and stability issues in decentralized optimization with compression.
method Proposes the first linearly convergent decentralized algorithm with compression.
result First consensus error bound for coupled dynamics of primal and dual updates.
We propose and analyze a new type of stochastic first order method: gradient descent with compressed iterates (GDCI). GDCI in each iteration first compresses the current iterate using a lossy randomized compression technique, and subsequently takes a gradient step. This method is a distillation of a key ingredient in t…
This paper compresses neural networks by permuting and quantizing weights.
problem Efficiently compressing large neural networks for resource-constrained platforms.
method Permuting and quantizing weights, connecting to rate-distortion theory, and using annealed quantization.
result Significant compression with minimal accuracy loss, e.g., 40-70% reduction in gap with uncompressed model.
Diffusion models improve image compression at low bit-rates.
problem Efficiently compressing images at very low bit-rates.
method Encoding into an embedding, using diffusion models to refine the embedding iteratively.
result Realistic reconstructions can be generated at extremely low bit-rates.
This work proves that large models can be compressed significantly without losing performance.
problem Achieving comparable performance with smaller models and less data.
method Developed a universal compression theory for neural networks and datasets.
result Proved that a generic permutation-invariant function can be compressed into a function of polylogarithmic size with vanishing error.
Proposes a new linearity-based neural network compression method.
problem Reduction of neural network model size while maintaining accuracy.
method Integrates linearity-based intuition with ReLU activation functions to merge layers.
result Achieves up to 75% reduction in model size without loss of accuracy.
Compressed sensing (CS) is a sampling theory that allows reconstruction of sparse (or compressible) signals from an incomplete number of measurements, using of a sensing mechanism implemented by an appropriate projection matrix. The CS theory is based on random Gaussian projection matrices, which satisfy recovery guara…
Detects handlebodies and mapping class extensions using bordered Floer homology.
problem Detecting handlebodies and mapping class extensions over compression bodies.
method Combining bordered Floer homology with ideas from Casson-Long and train tracks.
result Algorithm to detect mapping class extensions over any compression body.
Unified bounds linking compressibility, fractal dimensions, and mutual information.
problem Understanding generalization in stochastic learning algorithms.
method Rate-distortion theory applied to machine learning generalization.
result Unified bounds linking compressibility, fractal dimensions, and mutual information.
We show that model compression can improve the population risk of a pre-trained model, by studying the tradeoff between the decrease in the generalization error and the increase in the empirical risk with model compression. We first prove that model compression reduces an information-theoretic bound on the generalizati…
AdaBoost improves binary classification in robust one-bit compressed sensing with adversarial errors.
problem Binary classification in robust one-bit compressed sensing with adversarial errors.
method AdaBoost and max-ℓ1-margin-classifier approach, with convergence rates improved under certain feature conditions. result Improved convergence rates and explanation for harmless interpolating adversarial noise.
If M is a manifold with compressible boundary, we analyze essential disks in M, as well as incompressible, but not necessarily boundary incompressible, surfaces in M. We are most interested in the case where M is a handlebody or compression body. The analysis depends on a new normal surface theory. We hope the normal s…
The Information Plane theory predicts autoencoders do not compress input information.
problem Understanding the training dynamics of hidden layers in autoencoders.
method Derive a theoretical convergence for the Information Plane of autoencoders using a Gram-matrix based mutual information estimator.
result Ideal autoencoders with a large bottleneck layer size do not compress input information, while a small size causes compression only in the encoder layers.
Compression techniques for deep neural network models are becoming very important for the efficient execution of high-performance deep learning systems on edge-computing devices. The concept of model compression is also important for analyzing the generalization error of deep learning, known as the compression-based er…
Given Poincare spaces M and X, we study the possibility of compressing embeddings of M x I in X x I down to embeddings of M in X. This results in a new approach to embedding in the metastable range both in the smooth and Poincare duality categories.
Theoretical framework for neural network compression using sparsity norms.
problem Understanding and quantifying compressibility and accuracy trade-offs in neural networks.
method Using sparsity-sensitive ℓ_q-norm to characterize compressibility and developing adaptive pruning algorithms.
result Theoretical relationship between network sparsity and compressibility with controlled accuracy degradation.
Recurrent iterated function systems (RIFSs) are improvements of iterated function systems (IFSs) using elements of the theory of Marcovian stochastic processes which can produce more natural looking images. We construct new RIFSs consisting substantially of a vertical contraction factor function and nonlinear transform…
Paper proposes compressive ICA algorithms for ICA model.
problem Efficiently solving ICA model with reduced memory and computational complexity.
method Compressive learning approach to ICA model, proving existence of compressive ICA scheme, proposing two algorithms (IPG and ASD).
result Proposed algorithms achieve substantial memory gains over well-known ICA algorithms.
This work compresses heavy-tailed weight matrices for tighter generalization bounds.
problem Empirical evidence linking heavy-tailed weight matrices to test set accuracy but lack of formal relationship with generalization bounds.
method Utilized the compression framework to show that heavy-tailed matrices can be compressed, resulting in sparse weight matrices.
result Demonstrated a non-vacuous generalization bound for compressed networks with heavy-tailed weight matrices.
Sign-based algorithms (e.g. signSGD) have been proposed as a biased gradient compression technique to alleviate the communication bottleneck in training large neural networks across multiple workers. We show simple convex counter-examples where signSGD does not converge to the optimum. Further, even when it does conver…
Paper develops a decoder for sparse codes without encoder matrix, achieving optimal recovery.
problem Designing a decoder for sparse codes from linear measurements alone.
method Matrix factorization to recover encoder and sparse coding matrices from measurements.
result Decoder-Expander Based Factorisation recovers encoder and sparse coding matrix at optimal measurement rate with high probability.
Paper designs a lossy compression method for lossless prediction.
problem Ensuring high performance on predictive tasks with minimal data.
method Characterizes bit-rate requirements for invariant transformations, designs unsupervised objectives for neural compressors.
result Achieves substantial rate savings on ImageNet compared to JPEG without compromising classification performance.
IntSGD compresses SGD gradients without floats, converging as SGD.
problem Efficiently compressing stochastic gradients in distributed SGD.
method Adaptive integer compression of gradients, estimating scaling adaptively.
result IntSGD matches SGD's iteration complexity for convex and non-convex functions.
Studies on generalization performance of machine learning algorithms under the scope of information theory suggest that compressed representations can guarantee good generalization, inspiring many compression-based regularization methods. In this paper, we introduce REVE, a new regularization scheme. Noting that compre…
Model compresses event-like contexts using gated surprise signals.
problem Perceiving a dynamic world as organized events.
method Hierarchical, surprise-gated recurrent neural network architecture.
result Achieves best performance on multiple event processing tasks.
Modern deep learning models are often trained in parallel over a collection of distributed machines to reduce training time. In such settings, communication of model updates among machines becomes a significant performance bottleneck and various lossy update compression techniques have been proposed to alleviate this p…
We study the flow of information and the evolution of internal representations during deep neural network (DNN) training, aiming to demystify the compression aspect of the information bottleneck theory. The theory suggests that DNN training comprises a rapid fitting phase followed by a slower compression phase, in whic…
In the compressive learning theory, instead of solving a statistical learning problem from the input data, a so-called sketch is computed from the data prior to learning. The sketch has to capture enough information to solve the problem directly from it, allowing to discard the dataset from the memory. This is useful w…
New study reveals how heavy-tailed SGD dynamics lead to compressible neural networks.
problem Understanding why large neural networks can be compressed effectively.
method Linking SGD dynamics to compressibility properties of neural networks.
result Large step-size/batch-size ratios and overparametrization lead to heavy-tailed SGD dynamics, making networks compressible.
New SGD variant makes neural networks compressible without assumptions.
problem Improving neural network compressibility without strong assumptions.
method Introducing heavy-tailed noise to SGD iterates.
result Compressible outputs with high probability for any compression rate.
Efficiently compress overparameterized deep models by focusing on low-dimensional learning dynamics.
problem Overparameterized models increase computational and memory costs.
method Study of learning dynamics reveals updates occur within a low-dimensional subspace, leading to a compression algorithm.
result Compressed deep linear networks converge faster and yield smaller recovery errors.
This paper optimizes portfolio compression by reducing excess notional in market contracts.
problem Reduction of excess notional in market contracts to decrease systemic risk.
method Examines and solves portfolio compression problems using graph theory and algorithms.
result Developed a clearing algorithm and method to compute maximum volume conservative compression.
Investing is a compression problem, maximizing growth by minimizing divergence.
problem Maximizing long-term wealth and minimizing risk of ruin in investing.
method Decomposes investing into three terms: money, entropy, and divergence. Uses Kelly Criterion and universal portfolio theory.
result Investing can be seen as a compression problem, with optimal strategies minimizing divergence.
HOPE uses Hilbert space to deconstruct deep network representations.
problem Deconstructing learned representations in deep networks is challenging.
method Introduces Hilbert Operator for Progressive Encoding (HOPE) to deconstruct network weights.
result HOPE provides an unbiased approach to network compression and fine-tuning.
Optimized sampling scheme for compressed sensing combining randomness and determinism.
problem Improving compressed sensing performance with deterministic sampling.
method Optimized sampling scheme combining random and deterministic selection of rows.
result Measurable improvements in image compressed sensing for generative and sparse priors.
This paper connects Wasserstein distances to MMD norms for compressive statistical learning.
problem Comparing and controlling Wasserstein distances between probability distributions.
method Establishing conditions under which Wasserstein distances can be controlled by MMD norms.
result Introducing Wasserstein regularity for compressive statistical learning.
This paper presents a novel signal compression algorithm based on the Blaschke unwinding adaptive Fourier decomposition (AFD). The Blaschke unwinding AFD is a newly developed signal decomposition theory. It utilizes the Nevanlinna factorization and the maximal selection principle in each decomposition step, and achieve…
Recent years, many researches attempt to open the black box of deep neural networks and propose a various of theories to understand it. Among them, Information Bottleneck (IB) theory claims that there are two distinct phases consisting of fitting phase and compression phase in the course of training. This statement att…